[Bug target/48576] New: wrong code when accessing variables in a large stack frame

["akos.pasztory at gmail dot com" <gcc-bugzilla at gcc dot gnu dot org>]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48576

           Summary: wrong code when accessing variables in a large stack
                    frame
           Product: gcc
           Version: 4.5.2
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: akos.pasztory@gmail.com
            Target: arm-linux-gnueabi


Created attachment 23964
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=23964
preprocessed source

Tried with 4.5.2 (and 4.6.0), stage1 compiler:

../gcc-4.5.2/configure \
  --target=arm-linux-gnueabi \
  --disable-shared \
  --disable-bootstrap \
  --with-sysroot=/usr/arm-linux-gnueabi \
  --disable-threads \
  --disable-libmudflap \
  --disable-libssp \
  --with-arch=armv7-a --with-tune=cortex-a8 --with-float=hard --with-fpu=neon \
  --enable-languages=c \
  --with-newlib

When compiling the attached source with:

./cc1 -O1 -fgcse -fno-omit-frame-pointer jaj.i

gcc produces this code to do "arr[n] = strdup(str)":
...
    bl    strdup
    sub    r3, fp, #4096
    ldr    r3, [r3, #-132]  @ this kills r3
    ldr    r2, [r3, #-128]  @ and this doesn't know about it
    str    r0, [r2, r3, asl #2]
...

The problem is first visible in jaj.i.192r.ira when using -fdump-rtl-all. 
Omitting frame pointer "helps" but I think it's pure coincidence.  Apologies
about the messy test code but I was not able to reduce it more without changing
the output.